Workers inspect whether or not a crack exists on wall surfaces of structures such as bridges and tunnels. In order to reduce workers' inspection workload, there is a conventional technique for remotely capturing images of the wall surfaces of the structures and detecting a crack from the captured images. As examples of the conventional technique for detecting a crack from images, conventional techniques 1 and 2 are described below.
The conventional technique 1 (Japanese Laid-open Patent Publication No. 2012-98045) is to treat, as crack candidates, polygonal lines obtained by combining long straight lines and determine whether each of the crack candidates is a crack based on line segment likelihoods of local segments of the polygonal lines and polygonal line likelihoods obtained when the local segments are combined. In the conventional technique 1, similarities between line segments of the local segments and an approximate straight line are used as the line segment likelihoods, and line segment likelihoods of the line segments forming the polygonal lines are used as the polygonal line likelihoods.
The conventional technique 2 (Japanese Laid-open Patent Publication No. 2012-002531) is to determine that a crack candidate is not a crack if the crack candidate is a straight line extending in a horizontal or vertical direction. Other related-art techniques are Japanese Laid-open Patent Publications Nos. 2005-114671 and 2013-117409.
The aforementioned conventional techniques, however, have a problem in which the accuracy of determining whether or not a target object is a crack is low.
For example, a certain wall surface has a characteristic that is not a crack and is dirt, a groove of a structure, or the like and is similar to a crack. FIG. 16 is a diagram illustrating an example of a crack, dirt, and a groove of a structure. As illustrated in FIG. 16, a crack 1a, dirt 1b, and a groove 1c of a structure have similar characteristics. It is, therefore, difficult to accurately identify the crack.
In the conventional technique 1, it may be determined that non-linear dirt is not a crack, but it may be erroneously determined that a linear object such as a groove of a structure is a crack. In addition, in the conventional technique 1, if a target object is a crack and line segments forming a polygonal line are short, similarities between line segments of local segments and an approximate straight line are low and the detection of the crack may fail.
In the conventional technique 2, if a groove of a structure is not set to a line extending in a horizontal or vertical direction for imaging of a wall, the groove of the structure may be erroneously detected as a crack.
According to an aspect, an object of the present disclosure is to provide an image processing program, an image processing method, and an image processing device that may improve the accuracy of determining whether or not a target object is a crack.